[Feature] Wire Looper calls to the shared connector

[JiaoliangYu]

Related #3339

Purpose

Introduce the new Looper model-call path without removing the legacy transport yet, so the wiring can be reviewed independently from deletion.

• replace mutable per-client decision and Fusion metadata with typed, request-scoped ModelTarget and CallOptions
• add request-scoped headers to the shared model-runtime connector while keeping authorization precedence explicit
• route Looper calls through a non-retryable connector operation with bounded bodies and an explicit HTTP 200 success contract
• create one connector-backed Looper client per router generation and register it with the generation resource scope
• migrate Base, Confidence, Ratings, ReMoM, Workflows, Fusion, prompt selection, context recovery, and fusioneval call sites without changing their algorithm or response behavior

The legacy Looper HTTP branch remains in this PR only as a reviewable migration bridge. A focused follow-up removes it after every caller is on the typed path.

Test Plan

• go test -count=1 -race ./pkg/modelruntime/connector ./pkg/looper ./pkg/extproc
• go vet ./pkg/modelruntime/connector ./pkg/looper ./pkg/extproc
• make test-semantic-router PWD=/workspace on the stacked final tree
• make agent-ci-gate ENV=cpu CHANGED_FILES="..." on the stacked final tree

Test Result

All listed checks pass in rocm/atom-dev:latest on g64. The concurrency contract test overlaps two requests and verifies model, decision, iteration, Fusion depth, and authorization headers do not cross. The endpoint contract test verifies an empty operation path preserves the configured /v1/chat/completions endpoint exactly. The status-policy tests verify that the shared connector still accepts any 2xx by default while Looper rejects a streaming 204 response.

---

Semantic Router PR Checklist

• PR title begins with exactly one bracketed category
• The title uses one category only
• The PR links accepted issue [Refactor] Route Looper model calls through the shared connector #3339 with wg/mom-routing and one owner
• Commits in this PR are signed off with git commit -s
• Purpose, Test Plan, and Test Result reflect the actual scope and validation

[netlify]

✅ Deploy Preview for vllm-semantic-router ready!

Name	Link
🔨 Latest commit	9a742de
🔍 Latest deploy log	https://app.netlify.com/projects/vllm-semantic-router/deploys/6a9f657594654800085ad8ef
😎 Deploy Preview	https://deploy-preview-3508--vllm-semantic-router.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.
🤖 Make changes	Run an agent on this branch

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[JiaoliangYu]

The deletion-only follow-up is #3509. This keeps the connector wiring and the removal of the legacy transport independently reviewable.

[Xunzhuo]

Thanks for the careful connector migration. One response-contract regression remains: the shared connector accepts every 2xx response, while Looper’s existing path requires HTTP 200. Because the streaming parser accepts an empty body, a 204 now becomes a successful empty model response instead of an upstream-status error. Please preserve Looper’s 200-only contract and add a 204 regression test before removing the legacy path.

[JiaoliangYu]

Thanks for catching this — you are right that 204 could otherwise reach the streaming parser as a successful empty response. Fixed in 6ff1a1f9.

The connector operation now has an optional SuccessStatusCode: zero preserves the shared connector default of accepting any 2xx, while Looper explicitly requires 200 OK. I added both the connector policy contract test and a Looper streaming 204 regression test.

[Xunzhuo]

Thanks—SuccessStatusCode: http.StatusOK now preserves the Looper 200-only contract while the shared connector keeps its default 2xx behavior, and the 204 regression covers the original blocker. This branch is behind current main; please sync it, let the resulting exact-head checks finish, and re-request review.